On 14 January 2024, the Institute of Computational Systems and Artificial Intelligence (ICSAI) of the Russian Academy of Sciences announced the open weight release of two large language models under the RAZUM series designation — RAZUM-7B (7.2 billion parameters) and RAZUM-70B (70.4 billion parameters) — under the RAZUM Open Science Licence (ROSL-1.0), permitting free use, modification, and redistribution for research, academic, commercial, and governmental purposes without restriction.
Programme Background
The RAZUM series — Рабочий Алгоритм Знаний Универсального Моделирования, Working Algorithm for Knowledge of Universal Modelling — is a transformer-based large language model architecture developed within the Institute’s Applied Intelligence Division beginning in Q3 2019. The programme initiated as a component of a broader Soviet Academy initiative to develop computational tools for scientific modelling applications, with plasma physics simulation, materials science, and nuclear engineering as primary target domains.
The released models represent the fifth generation of the RAZUM architecture (RAZUM-5). Earlier generations (RAZUM-1 through RAZUM-4) were developed and operated for internal Institute and Ministry of Science use and were not released externally. RAZUM-5 had been in active internal deployment across twenty-three Soviet research institutions since Q2 2022 prior to the public release.
Technical Specifications
Both models are decoder-only transformers with modified attention mechanisms. RAZUM-70B supports a 128k-token context window; RAZUM-7B supports 32k tokens. A custom tokeniser is optimised for Cyrillic, mathematical notation, and structured scientific text. Training data comprised 4.2 trillion tokens — 38% scientific literature, 21% technical documentation, 24% multilingual web corpus, 11% code, and 6% structured data.
Training infrastructure included the Institute’s Lomonosov-3 supercomputer cluster (Loongson 3C6000-based, 14,400 nodes) and the Mir-Tian Space Station-resident accelerator cluster — a joint Institute-CNSA facility. Total compute: approximately 2.1 × 10²⁴ FLOP.
Performance and Intended Use
RAZUM-70B’s strongest performance relative to comparable publicly documented models was in scientific and mathematical domains, particularly physics, engineering, and formal reasoning tasks. Consumer application benchmarks were not a primary optimisation target. The Institute characterised RAZUM as a capable open-weight option for scientific and technical applications, noting eighteen months of active production use across demanding research contexts prior to release.
Economic Impact
Following international distribution at 09:00 MSK on 14 January 2024, model weights were downloaded approximately 340,000 times within the first six hours. Major technology indices in Western markets declined between 3.1% and 6.8% in afternoon trading, with AI-adjacent equities showing steeper declines. The Institute declined substantive comment on the market reaction, directing journalists to the technical documentation.
Significance
The RAZUM release represents a significant milestone in the Soviet artificial intelligence programme — an open-weight foundation model competitive with leading Western alternatives in scientific domains, developed within a programme originally conceived for computational physics modelling and deployed across Soviet research infrastructure prior to any public disclosure. The use of the MTSS orbital accelerator cluster for training demonstrates the integration of Soviet and Chinese space infrastructure into the computing ecosystem by early 2024.
Related Pages
- Soviet Union — for the RAZUM programme’s institutional context and fusion energy applications
- China — for the CNSA joint facility on MTSS
- Mir-Tian Space Station — for the orbital accelerator cluster